Does Capitalizations in Video Titles on YouTube Actually Drive Views? A case Study

Context

This analysis is aimed at investigating an apparent assumption when it comes to content creation, specifically on YouTube. Observation is enough to indicate that YouTubers commonly assume that capitalizing full words makes some sort of impact that generate more views.

To test it rigorously rather than anecdotally, I pulled real video data from a single, well-established YouTube channel (Daniel Cormier @DC_MMA)and ran the numbers. Daniel Cormier also known as DC was a former two-division champion in the UFC (Ultimate Fighting Championship) and is now a well respected commentator for the same organization. Alongside his duties as commentator for the UFC, he runs his own MMA-based YouTube channel. As implied, the analytical discourse and findings in this post are specific to the Daniel Cormier channel. Regardless, anything true about the DC channel can possibly be true for another similar channel.

The central question this analysis seeks to answer is: does capitalizations in a video title have a measurable impact on view count, after controlling for title length, video duration, and time since upload?

Data Collection and Manipulation

Data was retrieved directly from Daniel Cormier’s YouTube channel via the YouTube pubic Data API. For each video, the following variable were selected:

• View count (Dependent variable)
• Video title (to be used to extract):
  • Number of fully capitalized words, and
  • Total number of words
• Publish date/time (to be used to extract the number of days since video was uploaded)
• Video duration

Videos shorter than 240 seconds (4 minutes) were excluded from the dataset to filter out Shorts and very short clips, which tend to behave differently from standard long-form content in terms of view accumulation. After filtering, the working dataset contained 756 records.

Duration was originally returned by the API in ISO 8601 format (e.g., “PT11M28S”). It was converted into a numeric duration in seconds, then further converted into minutes for interpretability.

Variable Construction

Capitalization (CAP): Each title was split into individual words. Single-character tokens (e.g., stray “I,” “A,” or isolated punctuation) were removed entirely before counting, since including them would artificially inflate or distort the capitalization measure. From the remaining words, a raw count was taken of how many words were fully capitalized (e.g., “INSANE,” “UFC,” “REACTS”).

Title Length (WORDS): The total number of words remaining in each title after single-character removal, used as a control variable.

Duration (DUR): Video length in minutes.

Days Since Upload (DAYS): Calculated as the difference between the current date and each video’s publish date, intended to control for the fact that older videos have simply had more time to accumulate views.

Log Views (VIEWS): Raw view counts were heavily right-skewed, as is typical for this kind of data. To address this, view counts were log-transformed. See below.

Exploratory Findings

Before committing to a formal regression model, each predictor was plotted against log(views) to check whether a linear relationship, the core assumption of the planned regression approach, actually held in the data. See below.

It appears from visual inspection that capitalization and title length take on relatively few distinct integer values, causing points to cluster into vertical bands rather than a continuous spread. This discrete nature of the data limits how meaningfully linearity can be assessed for these two variables in the first place. Duration and ‘days since upload,’ which vary more continuously, show a weak but somewhat more discernible linear relationship with log(views), but still weak enough that none of the four predictors appear likely to generate meaningful predictive power in a regression analysis.

The Verdict

Proceeding to force a linear regression model onto this data would not have been the most honest or useful next step. Rather than manufacturing statistical significance through model tweaks or transformations chosen after the fact, the more defensible conclusion is that this channel’s data does not support a linear-regression-style relationship between these title/timing features and view performance. Specific to this channel, capitalization does not appear to help in generating views.

Discussion

Daniel Cormier’s channel is reactive and event-driven. His videos respond to specific, time-bound real-world events, a fight that happened on a particular date, a piece of breaking MMA news. The overwhelming majority of a video’s views arrive in the days immediately following its upload, driven by the topicality of the event itself. Once that news cycle passes, there’s little ongoing reason for new viewers to seek the video out. Nobody is searching for reactions to a fight from eight months ago.

With evergreen content channels such as tutorial videos for example, relevance doesn’t decay as it does for reactive videos. So these channels would be expected to show views continuing to accumulate steadily over time, producing a stronger, more clearly linear relationship between video age and view count, and thus more potential predictive value in a regression context.

Next Step

A reasonable follow-up is to repeat this same exploration process on a channel built around evergreen content. This will allow to to test the hypothesis raised above directly: that video age would show a positive relationship with views on a channel where content relevance doesn’t decay the way it does for event-driven content.